A Deep Learning Model for the Assessment of Filament Joining in Bi-2212 Composite Superconducting Wires
Matthew C. Jewell1 and Reed R. Oberg2
1Materials Science and Biomedical Engineering Department, University of Wisconsin-Eau Claire, Eau Claire, WI, 54702
2Department of Physics & Astronomy, University of Wisconsin-Eau Claire, Eau Claire, WI, 54702
Advances in high performance computing and open-source software to support machine learning have made new image analysis approaches accessible to a wide cross-section of researchers in the past decade. In this talk, we discuss the development of a semantic segmentation convolutional neural network that is aimed at solving complex imaging problems for superconducting materials. As an initial implementation of this model, we have applied our model to the assessment of filament joining in heat treated composite Bi2Sr2CaCu2O8-x (Bi2212) wires. Traditionally, this is a difficult task using quantitative image analysis techniques, since filament joining has a continuous range from lightly bridged to intimately conjoined. In this project we created a deep learning algorithm that classifies filaments from transverse cross-sectional wire images as either conjoined or individual in nature, based on training data provided by the user. The model can then subsequently classify filaments without human intervention. Overall, the model runs with an averaged accuracy of 94.1% for image pixels, and over 70% for filament categorization. With this model, more automated quality control processes for Bi-2212 wire production can be envisioned, and a more quantitative assessment of extent of conjoining can be provided to analyze the impact of bridging and agglomeration on the Jc and magnetization performance of the wire.
Acknowledgements: This research was supported by the U.S Department of Energy, Office of High Energy Physics, Award DE-SC0020984, and by the Blugold Center for High Performance Computing at UW-Eau Claire.
Zoom: https://lbnl.zoom.us/j/5104867866
Ian Pong, Mike Naus